Misfeat: Feature Selection for Subgroups With Mutual Information Estimation

Mutual information (MI) quantifies the dependence between features and target variables and is widely used in feature selection for downstream tasks. Due to the computational burden of exact MI computation, it is often estimated using statistical or learning-based methods. When data can be partitioned into subgroups, accurate MI estimation becomes even more challenging, as each subgroup may exhibit distinct featuretarget relationships. We introduce a framework that models feature-subgroup-target interactions as a multiplex graph and applies a heterogeneous graph neural network to propagate information between feature combinations both within and across subgroups. Our method efficiently identifies top- $K$ predictive feature subsets per subgroup while addressing key scalability challenges. Extensive empirical evaluation demonstrates the effectiveness and efficiency of our approach, offering, to the best of our knowledge, the first graph-based framework for subgroupaware MI estimation.